A machine learning-based regression technique for prediction of tropospheric phase delay on large-scale Sentinel-1 InSAR time series

Spatiotemporal variations in temperature, pressure, and relative humidity in the atmosphere produce the biggest source of error in InSAR data. Applying multi temporal interferometry (MTI) methods on the tropospherically corrected interferograms further improves the accuracy of velocity and displacement time-series. Interpolation of the external sources such as ERA-Interim model or the GNSS for tropospheric corrections is a big challenge, as we need to find a suitable function to predict the delay for the whole interferogram. Here, we propose a new technique based on machine learning (ML) Gaussian processes (GP) regression approach using the combination of small-baseline interferograms and GNSS derived zenith total delay (ZTD) values to mitigate tropospheric phase delay. The method facilitates the corrections, as we do not need to deal with finding a suitable function for interpolation of low resolution and/or sparsely distributed external observations. Applying our method on concatenated frames of Sentinel-1 images over Norway showed that the ML based method improves tropospheric corrections by 81% compared to 47% and 50% RMSE reduction gained by using ERA-Interim and GNSS only, respectively. Comparing the displacement time-series derived by small baseline interferograms corrected by our method with GNSS measurements showed overall RMSE of 5.2 mm.